 |
 |
This is what you're looking for: Chumbley & Friston. _False discovery rate revisited: FDR and topological inference using Gaussian random fields._ NeuroImage. 44(1):62-70 (2009) <http://www.sciencedirect.com/science?_ob=PublicationURL&_tockey=%23TOC%236968%232009%23999559998%23701157%23FLA%23&_cdi=6968&_pubType=J&view=c&_auth=y&_acct=C000057638&_version=1&_urlVersion=0&_userid=2503305&md5=933125709e50f7c83f8a328d6d9827f3> Hari Bharadwaj wrote: > Dear experts, > > As much as I appreciate that FDR control is a conceptual breakthrough in > the theory of multiple comparisons, I have what I think are fundamental > questions (which are probably already addressed in the literature): > > When I am searching a continuous space (such as a time-frequency map) for > significant effects, how do I interpret small patches (spanning say 2-3 > elements/pixels/voxels/vertices in each dimension) of significant effects > that show up? As discoveries, they are as valid as a big blob that might > show up. When the FDR control is done element-wise, it is clear that the > cluster-wise FDR could be high. Since we are actually going after clusters > or rather peaks in a functional map of some kind, what does voxelwise or > element-wise FDR control actually mean? It doesn't seem appropriate to > take the mass-univariate approach for a time-frequency map without any > consideration of the topology. What does it even mean to assign > activations to time-frequency elements when the underlying quantity is > inherently continuous? Am I missing something? FDR control seems to be > done routinely with functional imaging data. > > Thanks and Regards, > Hari > > >